Real-time voice agents are quickly becoming one of the most exciting things you can build in Python, but most tutorials skip straight to the demo and leave out everything that makes it actually work. Building your first voice agent often means slow responses, a bot that talks over itself, audio that cuts out mid-sentence, and no clear idea of what is broken or why. In this talk, we explore how Python developers can build a real-time voice agent from the ground up, without needing a machine learning background or a cloud budget that hurts. We draw on the experience of building and deploying a working voice agent in Python, one that listens, thinks, and responds in a way that feels natural rather than robotic. The talk begins with an overview of what makes real-time voice genuinely different from a chatbot: latency constraints, the audio pipeline, and the async patterns that hold everything together. We then walk through a clear, reproducible pipeline covering audio capture, speech-to-text, language model integration, and speech synthesis back to the user, with a focus on keeping each stage understandable and maintainable. A central part of the talk covers the parts most tutorials skip entirely: handling errors gracefully, streaming responses so the bot does not go silent while thinking, managing API keys safely, and making design decisions that turn a fragile demo into something you can actually hand to a user. Attendees will gain practical guidance on which tools are worth using, where the real complexity lives, and how to build something that works outside of a Colab notebook. This session emphasises conceptual clarity, working code examples, and real lessons from building and breaking this pipeline in practice. The audience will leave with a strong mental model of how real-time voice agents work, a reference architecture they can adapt to their own projects, and a concrete checklist for moving from demo to something people can actually talk to.

Predictive Maintenance (PdM) helps industries detect equipment faults before failure occurs, reducing downtime and maintenance costs. However, industrial datasets are often imbalanced because fault cases occur less frequently than normal operations, making accurate prediction difficult. This study benchmarks 4 ML classifiers (Logistic Regression, Random Forest, SVM, and XGBoost) for binary fault detection in rotating industrial machinery across 7,672 real-time sensor readings. To address class imbalance (≈90% normal, 10% fault), a stacking ensemble combining SVM, Random Forest, and XGBoost under a Logistic Regression meta-learner is proposed. The ensemble achieved 98.37% accuracy, F1 of 91.35, and AUC of 0.987, outperforming all individual classifiers. SHAP analysis identified temperature and pressure as dominant fault predictors, providing a deployable foundation for Industry 4.0 predictive maintenance in the Nigerian industry.